Click models (CMs) are designed to capture user interactions with information retrieval results, particularly focusing on click behavior in response to query-document pairs. A user’s clicks are influenced by several factors, including the relevance of the documents, the ranking of the results, and the user’s prior knowledge or biases. The click models aim to provide insights into user preferences, behavior patterns, and relevance assessment in the context of information retrieval. Various click models have been proposed, ranging from the probabilistic graphical models (PGM), to neural network based methods. The neural network based methods enhance probabilistic graphical models by increasing their expressive capacity and allowing for more flexible dependency modelling. However, challenges such as the data sparsity and cold-start problems remain. In this paper, we propose a novel retrieval-enhanced click model (RetrievalCM) for web search. First, we regard each query-document pair as a point in a well-defined latent space, and introduce effective methods for measuring distances between samples in this space, to fully exploit the historical behavior information. Second, after collecting the appropriate relevant samples, we model click prediction by combining current data together with historical behavioral data, addressing the data sparsity issue. Finally, we conduct extensive experiments to demonstrate the effectiveness of our approach.